Aishwarya SrinivasanSix diagrams, in article order
How the Model Context Protocol connects AI to tools and data: the problem it solves, the three pieces, the agent loop, the stack it complements, the safe way to use a server, and how to build your first one.
1 · From point-to-point chaos to one standard plug
Illustrates Sections 1–2 — why MCP exists (the drawer of chargers problem) and the USB-C fix it introduces.
Write the connector once — any MCP-compatible app (Claude, ChatGPT, Cursor, or your own agent) can use it: no rewrites, no special casing per platform.
Timeline: Anthropic released MCP in November 2024 and donated it to the Linux Foundation in December 2025. Downloads went from roughly 100,000 a month to 97 million a month in about eighteen months.
2 · Host, client, server — the three pieces
Illustrates Section 3 — how the three pieces map onto the phone, its USB-C port, and the accessory, plus the three things a server exposes.
Section 3
Host — the AI app · like your phone
Claude Desktop, Cursor, VS Code, ChatGPT — the application you actually talk to.
↓
MCP client — the USB-C port inside the host
Speaks the protocol and manages each connection. Invisible plumbing to you; the endpoint doing the talking to the protocol.
stdio or HTTP↓
MCP server — the accessory you plug in
A small program that wraps tools and data sources and exposes them in the standard MCP format. It can literally be about a hundred lines of Python on your laptop.
Tools
What the model can do — like giving a new hire software access so they can take actions.
Resources
What the model can read — like the company wiki for a new hire.
Prompts
Reusable templates for doing common tasks well — like your standard operating procedures.
When a client connects, it simply asks “What do you have?” and the server advertises its tools, resources, and prompts with descriptions and typed inputs — everything is discovered at runtime, nothing is hardcoded.
Example from the article: a filesystem server exposes read and write operations as tools and lets the model open files as resources.
3 · Where MCP sits: the agent loop
Illustrates Section 4 — an agent is a harness loop around the model; MCP is the company badge that gets the agent into every system it is allowed to touch.
Section 4
Model — the brainHarness — the workstation + workflowMCP — the company badge
1Userasks a question in plain language
↓
2Model · the braindecides it needs something external — pull rows from a database
↓
3Harness · the workstationroutes the intent through the MCP client to the right server
↓
4MCP clientcalls the matching tool on the server
↓
5MCP serverdoes the actual work and returns the rows
↓
6Harnessthe result flows back into the model’s context
↓
7Model · the brainkeeps reasoning with the new information
↓
↺ repeat until the task is complete — model decides, harness executes, MCP connects
This is why MCP took off together with agents: before MCP every tool meant teaching the agent a one-off interface; with MCP, any agent that speaks the protocol can use any server.
4 · MCP below function calling, above your APIs
Illustrates Section 5 — MCP replaces nothing you already use; it sits underneath the layers you have and standardizes the connection, which is why the plugin era ended.
Section 5
The stack MCP complements — nothing gets replaced
Function calling — a capability of the model
The model looks at the task, decides a tool is needed, and produces a structured request with the right arguments. “Should I call, and what should I say?”
↓
MCP — the protocol and transport
Carries that request to the tool and brings the result back. “How does the call actually reach the other end?” Function calling is dialing; MCP is the telephone network.
↓
Your API — does the actual work
A point-to-point connection to one service. Your APIs do not go anywhere on their own — MCP sits on top of them so every model can discover and call them the same way.
Why plugins lost (2023)
A plugin was proprietary: built for one platform, and getting the same capability anywhere else meant building it again. The old proprietary charger — one cable per brand.
How MCP flips it
One open protocol, and every platform implements the same port. Build the server once — it runs on Claude, ChatGPT, Cursor, VS Code, and whatever agent framework you use.
If someone tells you MCP replaces your APIs, they have misunderstood the stack: function calling is a model capability, MCP is the transport, and the API still does all the work.
5 · Local vs remote — and the gate before you plug in
Illustrates Sections 7–8 — the two flavors of server, the security gate most tutorials skip, and how a server actually gets used.
Sections 7–8
Start
You found a server to use — search first, because 10,000+ public servers exist and the integration you need probably already does.
↓
Security gate
Is it a verified publisher — or have you read the code yourself? Server code is usually small enough that you actually can.
no↓
Stop — do not run it
An MCP server is code with a direct line into your AI’s context and often your real accounts. Like plugging a USB stick found in a parking lot into your laptop.
yes↓
Local server — stdio
Runs on your machine and talks to the client over stdio (standard input/output between processes). Best for local files and experimentation.
Config: add the launch command that starts the server.
Remote server — HTTP
Hosted elsewhere; you connect over HTTP and typically authenticate with OAuth — the same way you sign in to any app with your Google account.
Config: add the server URL.
either path↓
Restart the app
Claude Desktop, Cursor, or VS Code picks up the new configuration entry.
↓
Handshake + discovery
The client handles the handshake and discovers the server’s tools automatically — no hardcoding.
↓
Use it
Talk to your AI normally — it calls the tools whenever it needs them.
Practitioner’s checklist — least privilege always (read-only credentials if it only reads), keep a human in the loop for writes, deletes, and messages; at team scale route servers through a central gateway or internal registry.
Recommended starting point: go local with something low-stakes like a filesystem server, watch how the model calls its tools, and build intuition there before scaling up.
6 · The fifteen-minute server: from function to live tool
Illustrates Section 9 — FastMCP turns a typed, documented function into a tool, the MCP Inspector validates it, and your client starts calling your code.
Section 9
1Pickone API or data source you use daily — Notion, a weather API, a read-only view of your own database
↓
2Writea normal Python function with type hints and a docstring
↓
3Decorateadd the FastMCP decorator — the SDK generates a fully described MCP tool, no protocol plumbing by hand
↓
4Verifytrigger the tool manually in the MCP Inspector before any model touches it
↓
5Registeradd the server to your Claude Desktop or Cursor configuration file
↓
6Liveyour AI calls your code over stdio — a real first server in about 15 minutes
scale later↓
Graduate only when needed
Move to HTTP with OAuth once more than one person needs the server. Most people overbuild their first server — don’t.
TypeScript developers: the official TypeScript SDK is just as solid as the Python one. Either way, you are giving function calling a standardized place to land.
Remember the one-liner from the article: MCP is not the intelligence part — it is the plumbing part. The teams winning with AI are the ones doing that plumbing well.
1 / 6
Use the tabs or Previous / Next to flip between diagrams; Play walkthrough highlights each stage in order.